Shadow-robust unsupervised flood mapping via GMM-enhanced generalized dual-polarization flood index and topography features

Abstract

Under cloudy and rainy conditions, Synthetic Aperture Radar (SAR) can provide essential data support for large-scale and high-timeliness flood monitoring. Although an optimal combination for Sentinel-1 VV-VH polarization data has been selected for flood mapping, its performance on complex terrains is not satisfactory, and it is unknown whether it can be further extended to generalize co- and cross-polarization data (not only VV-VH but also HH-HV). Therefore, we propose a shadow-robust unsupervised flood mapping method, called Topography and Dual-polarization Flood Mapping (TODFLOM). This method initially utilizes topography features to identify safe areas where floods are unlikely to occur in complex scenarios. Then, a Generalized Dual-Polarization Flood Index (GDPFI), based on the microwave scattering characteristics of complex inundation scenarios, is constructed to highlight flood features. Finally, GDPFI coupled with a Gaussian Mixture Model (GMM) is used for generating the flood map, enabling the method to effectively suppress mountain shadows in areas with slopes below 30°. The integration of topography features and GDPFI-GMM empowers TODFLOM to remove shadows impact with a 5° safety threshold. Experiments reveal that TODFLOM achieves an F1 score greater than 0.88 across all four flood datasets, outperforming other advanced methods in large-scale complex inundation scenarios.